---
license: cc-by-4.0
pretty_name: ARCADE
tags: [cad, cadquery, b-rep, step, point-cloud, mesh, code-generation]
configs:
- config_name: code
  data_files:
  - split: train
    path: code/train-*.parquet
  - split: validation
    path: code/validation-*.parquet
  default: true
- config_name: geometry
  data_files:
  - split: train
    path: geometry/train-*.parquet
  - split: validation
    path: geometry/validation-*.parquet
---
# ARCADE

CadQuery programs with their geometry, released with **StepCAD: Mesh-to-CAD Code Generation via LLM Policy and
Geometry-Guided Search** ([paper](https://arxiv.org/pdf/2610.03799), [code](https://github.com/ghadinehme/stepcad-v1),
[model](https://huggingface.co/DeCoDELab/StepCAD)).

**3,543,874 programs** (3,527,534 train, 16,340 validation) and **34,813,149 intermediate CAD states** (one per
`result = ...` operation), each with STEP, mesh and 10K-point cloud. Every program is in a canonical step-by-step
CadQuery format, re-executes to a valid single solid, and is unique (deduplicated by canonical program).

![statistics](arcade_stats.png)

| | ARCADE |
|---|---|
| programs | 3,543,874 |
| operations per program | 108.3 (68) |
| max operations | 6,115 |
| workplanes | 10.4 (9) |
| sketch primitives | 78.6 (38) |
| solid operations | 10.5 (9) |
| Booleans | 8.9 (7) |
| B-rep faces | 91.3 (63) |
| intermediate states | 34,813,149 |
| coverage: extrude | 94.7% |
| coverage: revolve | 28.3% |
| coverage: sweep | 18.6% |
| coverage: loft | 13.8% |
| coverage: fillet | 34.2% |
| coverage: chamfer | 25.6% |
| coverage: shell | 5.5% |
| coverage: mirror | 21.1% |

An operation is a modelling action: a workplane, a sketch primitive (line, arc, circle, rectangle, spline, ...), a
solid operation (extrude, revolve, sweep, loft, fillet, chamfer, shell, mirror) or a Boolean (union, cut,
intersect); selectors, `moveTo`, `close` etc. are not counted. Faces are B-rep faces of the final solid read back
from its STEP file. Coverage is the % of programs with at least one instance. Values are mean (median).
All numbers are in `release_metrics.json`.

## Layout

`code/` (default config, about 2 GB): one row per program.

| column | |
|---|---|
| `id` | `train_<16 hex>`, hash of the canonical program |
| `code` | CadQuery program |
| `num_ops` | number of `result = ...` operations |
| `face_count`, `volume` | B-rep faces and volume of the final solid |

`geometry/` (about 3.4 TB): one row per program, joined to `code/` on `id`. Files are named after their code shard
and first row.

| column | |
|---|---|
| `step` | final solid as a STEP file (bytes) |
| `step_faces`, `volume` | faces as read back from the STEP, volume |
| `mesh_vertices`, `mesh_faces` | triangle mesh: float32 (V, 3) and uint32 (F, 3), row-major bytes |
| `points`, `normals` | 10K surface points with normals of the final solid: float16 (10000, 3) bytes |
| `state_points`, `state_normals` | the same after every operation (list, one entry per state) |
| `state_faces` | B-rep face count after every operation |

```python
import numpy as np
from datasets import load_dataset
code = load_dataset("DeCoDELab/ARCADE", "code", split="train")
geom = load_dataset("DeCoDELab/ARCADE", "geometry", split="train", streaming=True)
r = next(iter(geom))
pts = np.frombuffer(r["points"], np.float16).reshape(-1, 3)
```

## License

CC-BY-4.0.